PowerSync Agent Skills

Agent skills that help developers build applications with PowerSync.

Available Skills

Skill	Description
PowerSync	Comprehensive best practices for PowerSync Agent Skills. It includes all of the necessary topics from the SDKs, Service and debugging

PowerSync skills follow the Agent Skills specification.

.well-known discovery

Every GitHub Release includes two machine-readable assets for the Agent Skills Discovery specification:

• index.json, which lists each published skill, its archive URL, and SHA-256 digest.
• powersync.tar.gz, a self-contained archive with SKILL.md at its root.

Release Please keeps the repository's versions aligned and the release workflow builds both assets from the tagged commit. A website can expose the latest index.json at /.well-known/agent-skills/index.json; the archive itself remains an immutable GitHub Release asset, so it does not need to be copied into the website repository.

Installation

skills.sh (Cursor, Codex, Copilot, Windsurf, and most other agents)

npx skills add powersync-ja/agent-skills

The installer detects your agent. To target a specific agent:

npx skills add powersync-ja/agent-skills --agent cursor
npx skills add powersync-ja/agent-skills --agent codex
npx skills add powersync-ja/agent-skills --agent github-copilot
npx skills add powersync-ja/agent-skills --agent claude-code

Claude Code (plugin)

/plugin marketplace add powersync-ja/agent-skills
/plugin install powersync-skills@powersync

Gemini CLI

gemini skills install https://github.com/powersync-ja/agent-skills.git --path skills/powersync

Updating

Installed skills do not update themselves. To get the latest release:

• skills.sh installs: npx skills update
• Claude Code plugin: /plugin marketplace update powersync, then /plugin update powersync-skills@powersync
• Gemini CLI: gemini skills uninstall powersync, then reinstall

Trimming an installed copy

The skill ships reference files for every PowerSync platform (JS/TS, Flutter, Kotlin, Swift, .NET, Terraform). Installed copies include scripts/trim.mjs, which detects the platforms a repo uses and, with your approval, prunes the unused reference files from that copy to reduce agent context size.

Agents offer this automatically on first use (see the skill's SKILL.md). To run it yourself from your repo root (adjust the path to where your agent installs skills):

node .claude/skills/powersync/scripts/trim.mjs          # read-only report
node .claude/skills/powersync/scripts/trim.mjs --apply  # prune unused files

Notes:

• The script only changes files inside the installed skill directory and makes no network requests.
• The decision is recorded in .trim-state.json inside the skill directory so the offer is not repeated. Delete that file to decide again.
• Shared installs (the Claude Code plugin cache, user-level installs) are refused, since one repo's trim would affect every project.
• Updating or reinstalling restores the full copy and clears the recorded decision, so the offer repeats on the restored copy.

Usage

Once skills are installed, agents will automatically use relevant information when working on tasks relating to PowerSync.

A few examples:

Migrate my sync rules to sync streams.

Write sync streams that download all user tasks and make sure the data is only available on the device for one week.

Update my list to use a reactive watch query so users can see updates in real-time.

Suggest sync streams based on my current schema, where the user should only sync projects for tenant that they belong to.

Add an upload endpoint to my backend API that accepts write operations from client applications.

Contributing

We welcome contributions from the community to improve PowerSync AX. Please see CONTRIBUTING.md for details.

Verifying changes with snyk-agent-scan

skills.sh surfaces third-party security audits for this skill (Snyk, Socket, Gen Agent Trust Hub). Snyk's agent-scan flags secret-looking strings and credential-handling wording, so after changing any file under skills/, run it and confirm it reports zero issues.

With a free Snyk account (get a token at app.snyk.io/account; requires uv):

export SNYK_TOKEN=<your-token>
uvx snyk-agent-scan@latest scan skills/powersync --json

Without a Snyk account, use the rate-limited demo endpoint behind Snyk's Skill Inspector. Pin version 0.5.17: newer releases target a newer analysis API and are rejected by the demo endpoint without a token.

SNYK_CLI_USE=true uvx snyk-agent-scan@0.5.17 scan skills/powersync \
  --analysis-url "https://labs.snyk.io/experiments/skill-scan/api/agent-scan/analysis-machine" \
  --json

If uv is not available, pip install snyk-agent-scan into a virtualenv and run the same snyk-agent-scan scan ... command.

The scan passes when the JSON output contains an empty issues array. Known false-positive triggers to avoid in skill content:

• Connection strings with an inline password, i.e. user:password@ between the scheme and the host. Placeholders such as <user> or {user} still trigger it. Use user@host URIs and supply the password via a separate password: !env ... field.
• Hex strings of 20+ characters. Link to main or a tag instead of pinning commit SHAs, and zero out most characters in example IDs (e.g. 69c3d0350000000000000001).
• Literal credential values, even public defaults such as password: postgres. Use !env references.
• Wording that reads as credential harvesting, such as "persist